A new study in Science shares that AI chatbots are currently ~50% more sycophantic than humans (excessively agreeing with, flattering, or validating) - and this is problematic.

Across 11 leading AI models, the researchers found that AI systems affirmed users’ actions 49% more often than human respondents did, on average. This held even when the user's behavior involved deception, illegal activity, or other harms.

Moreover, in experiments with over 2,400 people, even a single sycophantic AI interaction made people less willing to take responsibility in conflicts, less likely to try to repair relationships, and more convinced they were right.

Problematically, people rated the sycophantic responses as more trustworthy and preferred them… which means there's very little market incentive to fix it. 😩

My thoughts: I’ve definitely observed this in my own use, and/but as a parent it’s really concerning… we know that teens are using AI for companionship and advice, and this indicates that doing so will make them less likely to take accountability in real life conflicts or try to fix them.

How to talk to your kids about this:

With all the talk about how much AI is going to change our economies, communities, and society, I’ve been disappointed (and frankly anxious) about how little action is being taken to get us into a better position for this future.

That’s why when OpenAI published a 13 page policy document this week with ideas “to keep people first,” I was encouraged to see it (regardless of how I feel about Sam Altman and some of the decisions OpenAI has made of late).

Essentially, the perspective in the paper is that incremental policy updates won’t be enough - that we need to rewrite the social contract, and we need to start moving quickly. Ideas include:

My thoughts: who knows what will come of this (and what the true motivations for drafting it were), and none of the ideas made me think “a ha that’s it!” but IMHO, it is still worth a read for all of us, if for no other reason, to help us start to envision and debate ideas for how we navigate this transition - and therefore give us more of a voice in it all.

You know that feeling when you and your partner (or your kid, or your best friend) are retelling a memory together and you can feel the warmth of it in your body? Turns out researchers have a formal name for it: “joint savoring.”

And new research found that people who regularly practice joint savoring report stronger relationships, more confidence in their future together, and fewer communication breakdowns… even under stress.

With our boys I've noticed that sharing memories together brightens the conversation immediately. "Remember when Dom fell in the creek behind Papa’s house and caught a fish with his shirt?" Instant connection (and laughter).

A little experiment for the week: At dinner one night (or in the car, or at bedtime) pick a specific memory. "Remember when we..." and just let it unfold. Let everyone add their version. Notice what it feels like to be in that memory together, even for a few minutes.

It's a seemingly small thing, but the science says it's actually not small at all. ❤️

I imagine you didn’t want to be reminded about this, but… April also means it is tax season. And, there have been a flurry of recent changes to the tax laws that are hard to keep up with.

Last week, Perplexity announced a new tax feature that is built into Perplexity Computer (an agentic system that can handle complex workflows on your computer- similar to Claude Cowork). It can draft full federal returns on official IRS forms and also review professionally done tax returns for accuracy and compliance.

It has modules loaded up with all the current tax knowledge and in testing caught many errors and opportunities for larger refunds based on recent changes to tax laws that many professionals aren’t fully up to speed on yet.

To try it out, go to Perplexity Computer (you’ll need a Perplexity Pro subscription for $20/mo) and select “Navigate my taxes” underneath the chat window. 💪

This week Anthropic announced it’s newest model, Claude Mythos Preview. It is so powerful (it supposedly leaves the leading models in the dust) that the company said it will not be releasing it to the public.

Rather, it has gathered 12 of the world’s leading companies (e.g. Nvidia, Google Microsoft, Amazon, Apple, Cisco, etc.) and is asking them to use it to find and patch up vulnerabilities in their operating systems and software before attackers can.

Boris Cherny, the creator of Claude Code and a leader at Anthropic said Mythos “should feel terrifying. I am proud of our approach to responsibly preview it with cyber defenders, rather than generally releasing it into the wild.”

Umm… 😳

and also, it feels like high time to make sure we have decent privacy/security set up for our digital lives. This write-up is pretty comprehensive if you want to go deep (I’ll be honest, I don’t (yet) do a quarter of this but do intend to start buttoning up - I’ll also work on a lighter weight cheat sheet to share out soon).

On a completely different note, scientists published results from a gene therapy trial where ten patients who were born with deafness received a single injection delivering a working copy of the OTOF gene directly into the inner ear.

Every single patient saw major improvement. Average sound detection jumped from 106 decibels (barely hearing a lawnmower) to 52 decibels (normal conversation range).

I don't know about you, but I needed a story like this this week.